Lung cancer remains the leading cause of cancer-related mortality worldwide, with non–small cell lung cancer (NSCLC) accounting for ~85% of cases (1). Despite advances in targeted therapy and immunotherapy, many patients, particularly those with advanced-stage disease, continue to experience poor outcomes, underscoring the importance of early, accurate treatment response assessment to guide timely therapeutic decisions and avoid ineffective toxicity (2, 3).

Radiologic response assessment in routine practice relies primarily on standardized criteria such as RECIST 1.1, applied by expert radiologists across serial imaging studies (4). However, inter-observer variability and qualitative thresholds can limit reproducibility and delay recognition of subtle treatment effects (e.g., inflammatory changes, pseudoprogression), potentially leading to under or over estimating efficacy (5, 6).

Artificial intelligence (AI) systems spanning radiomics pipelines and deep learning architectures can quantify high-dimensional image features and temporal changes beyond human perception, promising earlier and potentially more objective prediction of treatment response (7–10). Early studies in thoracic oncology suggest AI may match or exceed radiologists for specific tasks (e.g., response prediction on CT or PET/CT). Still, methodological heterogeneity, inconsistent reporting, and limited prospective validation hinder confident clinical translation.

Prior reviews have mainly focused on diagnosis or broad oncologic use cases rather than the comparative performance of AI versus radiologists specifically for treatment response prediction in lung cancer. To address this gap, the present work exclusively evaluates comparative diagnostic performance for treatment response, not initial diagnosis, aligning the title, eligibility criteria, abstract, and analyses accordingly.

We conducted a PRISMA-guided systematic review and meta-analysis to synthesize pooled sensitivity, specificity, and accuracy for AI systems versus radiologists in predicting lung cancer treatment response, with prespecified subgroup analyses (by imaging modality and clinical context) and comprehensive assessment of heterogeneity (I²), threshold effects, sensitivity analyses (leave-one-out), and publication bias (funnel/Egger). We also expand on interpretability (e.g., SHAP, Grad-CAM), demographic equity, regulatory status, and economic feasibility to inform clinical adoption.

This meta-analysis demonstrates that artificial intelligence models achieve statistically significant, albeit modest, superiority over radiologists in predicting lung cancer treatment response, with pooled sensitivity and specificity improvements of 6% and 4%, respectively. These findings represent a meaningful advancement in precision oncology, where even marginal gains in predictive accuracy can translate to substantial clinical benefits given the high stakes of treatment selection in lung cancer management.

The observed performance advantage was most pronounced in PET and CT imaging, modalities that form the cornerstone of treatment response evaluation in current clinical practice (19). This modality-specific superiority likely reflects the inherent characteristics of these imaging techniques and their alignment with AI pattern recognition capabilities. PET imaging’s quantitative metabolic information provides rich data for machine learning algorithms to identify subtle changes in tumor glucose uptake that may precede morphological changes detectable by human interpretation. Similarly, CT’s high spatial resolution and standardized acquisition protocols create consistent datasets that facilitate robust AI model training and validation. In contrast, MRI-only analyses did not show a statistically significant AI advantage in our pooled results, underscoring the need for modality-specific validation.

The predominance of convolutional neural network (CNN)-based architectures among the evaluated AI systems aligns with the established superiority of deep learning approaches in medical image analysis (20). CNNs excel at hierarchical feature extraction, automatically identifying complex patterns across multiple scales that may escape human perception. However, the absence of regulatory approval for any evaluated systems highlights a critical gap between research innovation and clinical translation (21). This regulatory void raises fundamental questions about quality assurance, liability frameworks, and reimbursement mechanisms that must be addressed before widespread clinical adoption.

The clinical implications of AI-assisted treatment response prediction extend beyond diagnostic accuracy to encompass broader aspects of personalized cancer care. Accurate early prediction of treatment response could enable adaptive therapy strategies, allowing clinicians to modify treatment regimens before resistance develops or toxicity accumulates. This paradigm shift from reactive to proactive treatment management represents a fundamental evolution in oncological practice, potentially improving both survival outcomes and quality of life for patients with lung cancer.

However, integrating AI into clinical decision-making workflows presents complex challenges that transcend technical performance metrics. The “black box” nature of deep learning models creates interpretability barriers that may impede clinician acceptance and patient trust. While several studies incorporated explainable AI methods such as Gradient-weighted Class Activation Mapping (Grad-CAM), none systematically evaluated their impact on clinical decision-making or patient outcomes (22). Complementary frameworks such as SHAP can also clarify feature contributions (e.g., tumor texture, volume, or metabolic intensity), and embedding these explanations into reporting/PACS could facilitate human AI collaboration in practice (23).

The demographic homogeneity observed across the included studies raises essential questions about AI generalizability and health equity (24). The underrepresentation of diverse patient populations in training datasets may perpetuate healthcare disparities, as AI models may perform suboptimally in underrepresented groups which is constant in other studies as well. This concern is particularly relevant in lung cancer, where significant racial and socioeconomic disparities in outcomes already exist. Ensuring equitable AI performance across diverse populations will require deliberate efforts to include representative datasets and conduct subgroup analyses during model development and validation.

Economic considerations represent another critical dimension of AI implementation that remains largely unexplored in the current literature (25). While the direct costs of AI systems are substantial, encompassing software licensing, hardware infrastructure, and integration expenses, the potential for cost savings through improved treatment selection and reduced unnecessary interventions may justify these investments. The economic burden of lung cancer treatment, exceeding $21 billion annually in the United States, suggests that even modest improvements in treatment response prediction could yield significant healthcare savings through optimized resource allocation and reduced treatment failures. Formal cost-effectiveness analyses were not reported across the included studies and should be incorporated prospectively.

The heterogeneity observed across studies in AI architectures, training methodologies, and validation approaches reflects the nascent state of the field and the absence of standardized evaluation frameworks. This variability complicates direct comparisons between AI systems and may contribute to the moderate statistical heterogeneity observed in our meta-analysis (I² values in the mild range). We found no evidence of a threshold effect based on Spearman’s correlation, and leave-one-out sensitivity analyses did not materially change the pooled estimates, supporting the robustness of the findings. Funnel plot symmetry and Egger’s test were consistent with a low likelihood of publication bias (p = 0.21), although selective reporting cannot be entirely excluded. Potential overlap of patient cohorts from the same institutions could not be wholly resolved and may have modestly inflated pooled sample size estimates. Finally, the protocol for this review was not prospectively registered, which introduces some risk of reporting bias despite our PRISMA-guided methods.

The temporal dynamics of treatment response prediction present additional complexity that warrants consideration. Traditional imaging-based response assessment typically occurs at predetermined intervals, often weeks to months after treatment initiation. AI models capable of earlier response prediction could enable more timely treatment modifications, potentially improving outcomes while minimizing exposure to ineffective therapies. However, the optimal timing for AI-assisted response prediction and its integration with existing clinical protocols requires further investigation.

Future research should prioritize prospective, multicenter validation studies that address the limitations identified in this meta-analysis. Such studies should incorporate diverse patient populations, standardized imaging protocols, and comprehensive economic evaluations to provide robust evidence for clinical implementation. Additionally, the development of hybrid human-AI decision support systems that leverage the complementary strengths of both approaches may offer superior performance compared to either modality alone. The findings of this meta-analysis position AI as a promising adjunct to radiologist interpretation rather than a replacement technology (26). The modest but consistent performance improvements observed across multiple studies suggest that AI can enhance human expertise while preserving the critical role of clinical judgment in cancer care. This collaborative approach may optimize the benefits of both human experience and machine precision, ultimately improving patient outcomes through more accurate and timely treatment response prediction.

This meta-analysis provides one of the most comprehensive evaluations comparing artificial intelligence (AI) and radiologist performance in predicting lung cancer treatment response. Key strengths include a large pooled sample size of 6,615 patients across 11 independent studies, which confers strong statistical power. The methodology adhered to PRISMA 2020 guidelines, incorporating a multi-database search strategy (PubMed, Embase, Scopus, Web of Science, and Cochrane Library) with clearly defined inclusion criteria. Multiple diagnostic performance metrics (sensitivity, specificity, accuracy, and risk difference) were analyzed using standardized statistical models in RevMan, with transparent visualization through forest plots. Independent dual-reviewer screening and data extraction enhanced methodological reliability, with inter-rater agreement quantified (Cohen’s κ = 0.84). In addition, subgroup analyses by imaging modality (CT, PET, MRI) and clinical context (e.g., advanced disease, EGFR-targeted therapy) provided clinically relevant insights.

Several limitations should also be acknowledged. First, the predominance of retrospective designs (10 of 11 studies) increases susceptibility to selection bias and limits external validity. Second, potential population overlap between studies from the same institutions could not be fully resolved due to limited reporting, potentially inflating pooled sample sizes. Third, heterogeneity in imaging modalities, AI architectures, and “treatment response” definitions may affect comparability and introduce threshold effects. Fourth, incomplete demographic reporting across studies limits assessment of equity and generalizability to diverse patient populations. Fifth, none of the included AI systems had regulatory clearance, restricting immediate clinical applicability. Sixth, cost-effectiveness analyses were absent, despite economic feasibility being critical for adoption in resource-limited healthcare systems. Seventh, interpretability tools were rarely incorporated, with limited application of explainable AI approaches such as Grad-CAM or SHAP. Finally, while efforts were made to minimize reporting bias, publication bias remains possible due to underreporting negative or neutral findings.

This meta-analysis demonstrates that AI models achieve modest but statistically significant superiority over radiologists in predicting lung cancer treatment response, with pooled gains in sensitivity and accuracy, particularly in CT and PET imaging and in subgroups with advanced disease or EGFR-targeted therapy. These findings highlight AI’s potential as a valuable adjunct to human expertise, capable of enhancing diagnostic precision while maintaining comparable specificity and safety.

However, the predominance of retrospective studies, incomplete demographic reporting, heterogeneity in AI architectures, and lack of regulatory clearance limit the immediate generalizability of these results. Furthermore, few studies have incorporated explainable AI methods or assessed cost-effectiveness, which are factors critical for clinical adoption.

Future research should prioritize large-scale, multicenter, prospective trials that evaluate AI in real-world workflows, incorporate transparent and interpretable algorithms, assess equity across diverse populations, and address economic feasibility. AI integration should be conceptualized as augmenting, not replacing, radiologist judgment, fostering a multidisciplinary, patient-centered approach to lung cancer care. Additionally, the MRI-only subgroup did not show a significant AI advantage, and publication bias appeared unlikely based on Egger’s test (p = 0.21).